Some of these indexing implementations use a cache the length of the entire column (and not just the subset), which does indeed seem suboptimal. That’s likely where this is coming from. Finding a way to limit that to the size of the index would be great.